QSU Seminar Reliability and Validity: Design and Analytic Approaches Practical Considerations Rita Popat, PhD Dept of Health Research & Policy Division of Epidemiology rpopat@stanford.edu What do we want to know about the measurements? Why? Dependent variable (outcome) Independent variable (risk factor or predictor) JAMA. 2004; 292:1188-1194 2 What are other possible explanations for not detecting an association? JAMA. 2004;291:1978-1986 4 Outline • Definitions: Measurement error, reliability, validity • Why should we care about measurement error? • Effects of measurement error on Study Validity (Categorical Exposures) • Effects of measurement error on Study Validity (Continuous Exposures) • Measures (or indices) for reliability and validity Measurement error • For an individual, measurement error is the difference between his/her observed and true measurement. • Measurement error can occur in dependent (outcome) or independent (predictor or exposure) variables • For categorical variables, measurement error is referred to as misclassification • Measurement error is an important source of bias that can threaten internal validity of a study 6 Reliability (aka reproducibility, consistency) • Reliability is the extent to which repeated measurement of a stable phenomenon- by the same person or different people and instruments, at different times and places- obtain similar results. • A precise measurement is reproducible, that is, has the same (or nearly the same) value each time it is measured. • The higher the reliability, the greater the statistical power for a fixed sample size • Reliability is affected by random error 7 Validity or Accuracy • The accuracy of a variable is the degree to which it actually represents what it is intended to represent • That is: The extent to which the measurement represents the true value of the attribute being assessed. 8 Precise (Reliable) and Accurate (Valid) measurements are key to minimizing measurement error Precision, no accuracy No precision, no accuracy Accuracy, low precision Precision and accuracy 9 Measurement error in Categorical Variables • Referred to as Misclassification and could be in the • Outcome variables, or • Exposure variables • How do we know misclassification exists? • When method used for classifying exposure lacks accuracy Assessment of Accuracy Criterion validity (compare against a reference or gold standard) True classification Imperfect Classification Present + Present - Sensitivity = a / (a+c) Specificity = d / (b+d) Present + a TP c FN a+c Absent b FP d TN b+d a+b c+d False negative = c / (a+c) False positive = b / (b+d) 11 Misclassification of exposure Exposure + Cases (outcome +) a Controls (outcome -) b Exposure - c d • Non-differential • Differential 12 Misclassification of exposure True exposure Cases Controls Exposure + 50 20 Exposure - 50 80 (50)(80) OR 4.0 (20)(50) Reported exposure:90% sensitivity & 80% specificity in cases & controls Cases Controls Exposure + 55 34 Exposure - 45 66 (55)(66) OR 2.4 (34)(45) Attenuation of true association due to misclassification of exposure 13 Misclassification of the exposure Exposure + Cases (outcome +) a Controls (outcome -) b Exposure - c d • Non-differential misclassification occurs when the degree of misclassification of exposure is independent of outcome/disease status • Tends to bias the association toward the null • Occurs when the sensitivity and specificity of the classification of exposure are same for those with and without the outcome but less than 100% 14 Underestimation of a relative risk or odds ratio for… 0 0 Observed value True value 1 A. Risk factor 2 1 B. Protective factor 2 Bias toward the null hypothesis True value Bias toward the null hypothesis Observed value Modified from Greenberg. Fig 10-4, chapter 10 Misclassification of the exposure True exposure Cases Controls Exposure + 50 20 Exposure - 50 80 Reported exposure: Cases (50)(80) OR 4.0 (20)(50) Cases - 96% sensitivity and 100% specificity Controls- 70% sensitivity and 100% specificity Controls Exposure + 48 14 Exposure - 52 86 (48)(86) OR 5 .7 (52)(14) 16 Misclassification of the exposure True exposure Cases Controls Exposure + 50 20 Exposure - 50 80 Reported exposure: Cases (50)(80) OR 4.0 (20)(50) Cases - 96% sensitivity and 100% specificity Controls- 70% sensitivity and 80% specificity Controls Exposure + 48 30 Exposure - 52 70 (48)(70) OR 2 .1 (52)(30) 17 Misclassification of the exposure Exposure + Cases (outcome +) a Controls (outcome -) b Exposure - c d Differential misclassification occurs when the degree of misclassification differs between the groups being compared. • May bias the association either toward or away from the null hypothesis • Occurs when the sensitivity and specificity of the classification of exposure differ for those with and without the outcome 18 Overestimation of a relative risk or odds ratio for… 0 Bias away from the null hypothesis True value Observed value 1 A. Risk factor 2 Bias away from the null hypothesis Observed value 0 True value 1 B. Protective factor Modified from Greenberg. Fig 10-4, chapter 10 2 Hormone therapy Cases Controls Index Index Proxy (~25%) Proxy (~25%) Never Former Current accuracy? Pharmacy database Summary so far…. • Misclassification of exposure is an important source of bias • Good to know something about the validity of measurement for exposure classification before the study begins • Almost impossible to avoid misclassification, but try to avoid differential misclassification • If the study has already been conducted, develop analytic strategies that explore exposure misclassification as a possible explanation of the observed results (especially for a “primary” exposure of interest) Measurement error in Continuous Variables • Physiologic measures (SBP, BMI) • Biomarkers (hormone levels, lipids) • Nutrients • Environmental exposures • Outcome measures (QOL, function) Model of measurement error Measurement Theory: Example contd.. Measurement error Ti Ti + b Ti + b + Ei _________ b _______ ______________ E _______________ + systematic error in + additional "random error" for subject i instrument (bias) EXAMPLE: One m easured diast olic blood pressure (DBP) as indicat or of 2-year average DBP. BD cuff miscalibrated -measures everyone's diastolic BP as 10 mm Hg less + randomness in BP cuff mechanics + subject i's 10 mmHg increase over 2-year average + subject intimidated - diastolic BP 20 mmHg higher than u + misreading by interviewer + random fluctuations in current BP 24 Measurement theory 25 Validity of X… Measurement error Differential Measurement error OR Differential Measurement error OR Differential Bias Non- Differential Measurement error The effects of non-differential measurement error on the odds ratio. ORT is the true odds ratio for exposure versus reference level r. ORO is the observable odds ratio for exposure versus reference level r. Effects of non-differential measurement error 2 1 / XT O RRT RR 0.821/ 0.62 0.73 Summary so far…. • Measurement error is an important source of bias • Good to know something about the validity of measurement for exposure before the study begins • Almost impossible to avoid misclassification, but try to avoid differential misclassification! • Non-differential measurement error will attenuate the results towards the null, resulting in loss of power for a fixed sample size • This should be taken into account when estimating sample size during the planning stage and • Interpretation of results and determining internal validity of a study So why should we evaluate reliability and validity of measurements? • If it precedes the actual study, it tells us whether the instrument/method we are using is reliable and valid • This information can help us run sensitivity analysis or correct for the measurement error in the variables after the study has been completed Outline • Definitions: Measurement error, reliability, validity • Why should we care about measurement error? • Effects of measurement error on Study Validity (Categorical Exposures) • Effects of measurement error on Study Validity (Continuous Exposures) • Measures (or indices) for reliability and validity Choice of reliability and validity measures depend on type of variable . . . Type of Variable Reliability Measure(s) Validity Measure(s) Dichotomous Kappa Ordinal weighted kappa ICC* Continuous ICC * Bland Altman Plots sensitivity, specificity misclassification matrix Pearson correlation (see note) Bland-Altman Plots *ICC – intraclass correlation coefficient Note: in inter-method reliability studies, inferences about validity can be made from coefficients of reproducibility (such as the Pearson’s correlation ) 38 Assessing Accuracy (Validity) of continuous measures • Bias: difference between the mean value as measured x and the mean of the true values X • So bias = x – X • Standardized bias = x X SDX • Bland-Altman plots 39 Bland and Altman plots • Take two measurements (different methods or instrument) on the same subject • For each subject, plot the difference b/w the two measures (y axis) vs. the mean of the two measures • We expect the mean difference to be 0 • We expect 95% of the differences to be within 2 standard deviations (SD) Yoong et al. BMC Medical Research Methodology 2013, 13:38 Yoong et al. BMC Medical Research Methodology 2013, 13:38 Suppose there is no gold standard, then how do we evaluate validity? ….. We make inferences from inter-method reliability studies! Note: will not be able to estimate bias when the two measures are based on different scales Inferences about validity from inter-method reliability studies • Suppose two different methods (instruments) are used to measure the same continuous exposure. Let X1 denote the measure of interest (i.e., the one to be used to measure the exposure in the study) and X2 is the comparison measure • We have the reliability coefficient x x 1 2 • However, we are actually interested in the validity coefficient: Tx 1 • Example: Is self-reported physical activity valid? Compare it to the 4-week diary. 44 Relationship of Reliability to Validity Errors of X1 and X2 are: Relationship b/w reliability and validity Usual application 1. Uncorrelated and both measures are equally precise Tx Tx x x Intramethod study 2. Uncorrelated , X2 is more precise than X1 x x Tx x x Intermethod study 3. Uncorrelated , X1 is more precise than X2 4. Correlated errors and both measures are equally precise 1 2 1 2 1 2 1 1 2 Tx x x 1 1 2 Tx x x 1 Intermethod study Intramethod study Intermethod study 1 2 Take home message: In most situations the square root of the reliability coefficient can provide an upper limit to the validity coefficient 45 Inferences about validity from inter-method reliability studies • In our example, X1 is measure of interest (i.e., the one to be used to measure the exposure in the study: self-reported activity) and X2 is the comparison measure (4-wk diaries) • We have the reliability coefficient x x 1 2 = 0.79 • Errors in X1 and X2 are likely to be uncorrelated and X2 is more precise than X1, so 0.79 < Tx 1 < 0.89 • So, self-reported activity appears to be a valid measure 46 Summary of Inferences From Reliability to Validity • Reliability studies are used to interpret validity of x. • Reliability is necessary for validity (instrument cannot be valid if it is not reproducible). • Reliability is not sufficient for validity - repetition of test may yield same result because both X1 and X2 measure some systematic error (i.e., errors are correlated). • Reliability can only give an upper limit on validity. If the upper limit is low, then the instrument is not valid. • An estimate of reliability (or validity) depends on the sample (i.e., may vary by age, gender, etc.) 47 Reliability of continuously distributed variables • Pearson product-moment correlation? • Spearman rank correlation? 48 But…does correlation tell you about relationship or agreement? 185 180 175 Measure 2 Measure 1 Measure 2 150 155 155 158 160 165 163 170 170 176 174 184 190 170 165 160 155 150 145 150 155 160 165 170 175 180 Measure 1 Pearson’s Correlation coefficient=0.99 Is this measure reliable? 49 Reliability of continuously distributed variables • Other methods generally preferred for intra or interobserver reliability when same method/instrument is used - Intraclass correlation coefficients (ICC): is calculated using variance estimates obtained through an analysis of variance (ANOVA) - Bland-Altman plots • Correlation coefficient useful for inter-method reliability to make inferences about validity (especially when the measurement scale differs for the two methods) 50 Intraclass Correlation Coefficient (ICC) Total variance ICC = Within person Between person Between person variance Total variance • If within-person variance is very high, then measurement error can "overwhelm" the measurement of between person differences. • If between-person differences are obscured by measurement error, it becomes difficult to demonstrate a correlation between the imperfectly measured characteristic and any other variable of interest. • ICC is computed using ANOVA 51 ANalysis Of Variance (ANOVA) in a reliability study In a reliability study, we are not studying associations b/w predictors and outcome, so we will express the overall variability in the measurement as a function of between-subjects and within-subjects variability SST = SSB + SSW • So let’s consider a test-retest reliability study, where multiple measurements are taken for each subject 52 Total Variation SST ( X 11 X ) ( X 12 X ) ... ( X k n X ) 2 2 2 n Response, X X Subject 1 1 Group Subject Group22 Subject33 Group 53 Between-Subject Variation SSB k1 ( X1 X )2 k2 ( X 2 X )2 ... kn ( X n X )2 Where: k1= number of measurements taken on subject 1 Response, X X3 X2 X1 Subject Group 11 Subject22 Group X Subject Group 3 3 54 Within-Subject Variation (continued) SSW ( X 11 X 1 ) ( X 12 X 2 ) ... ( X k n X n ) 2 2 2 n Response, X X 12 X 13 X3 X11 X1 Group 11 Subject Group Subject22 X2 Group 3 3 Subject 55 One way analysis of variance for computation of ICC: testretest study Source of variance Sum of squares (SS) Degrees of freedom (df) Mean square (MS=SS/df) n-1 BMS ij X i n (k –1) WMS ij X nk - 1 k X i X 2 Between subjects Within subjects (random error) Total i X i j X i 2 2 j Here, each subject is a group. k=# times measure is repeated ICC ˆ x BMS WMS BMS ( k 1)WMS Interpretation of ICC ICC ˆ x BMS WMS BMS ( k 1)WMS • If within-person variance is very high, then measurement error can "overwhelm" the measurement of between person differences. • If between-person differences are obscured by measurement error, it becomes difficult to demonstrate a correlation between the imperfectly measured characteristic and any other variable of interest. 57 Interpretation of ICC • The ICC ranges b/w 0 and 1 and is a measure of reliability adjusted for chance agreement • An ICC of 1 is obtained when there is perfect agreement and in general a higher ICC is obtained when the within-subject error (i.e., random error) is small. • Hence, ICC=1 only when there is exact agreement between measures (i.e., Xi1=Xi2=...Xik for each subject). • Generally, ICCs greater than 0.7 are considered to indicate good reliability. 58 Two-way fixed effects ANOVA for computation of ICC (inter-rater reliability) Source of variance Mean square Sum of squares (SS) Degrees of freedom (df) (MS=SS/df) k X i X n-1 SMS Between measures n X j X k-1 MMS Within subjects (random error) X 2 Between subjects i 2 j i Total ij X 2 EMS j by subtraction ˆ x (n-1)(k-1) nk - 1 n( SMS EMS ) nSMS ( k 1) MMS ( n 1)( k 1) EMS 59 Measuring reliability of categorical variables • Percent agreement or concordance rate • Kappa statistic 60 Reliability of categorical variables • Concordance rate is the proportion of observations on which the two observers agree • Example: Agreement matrix for radiologists reading mammography for breast cancer Radiologist A Radiologist B Yes + No - Yes + No - a b a+b c d c+d a+c b+d Overall % agreement = (a+d) / (a+b+c+d) 61 Concordance rates: limitations • Considerable agreement could be expected by chance alone. • Misleading when the observations are not evenly distributed among the categories (i.e., when the proportion “abnormal” on a dichotomous test is substantially different from 50%) So, what reliability measures should we use? 62 Kappa • Kappa is another measurement of reliability • Kappa measures the extent of agreement beyond that would be expected by chance alone • Can be used for binary or variables with >2 levels 63 Cohen’s Kappa ( ): some notation • A reliability study in which n subjects have each been measured twice where each measure is a nominal variable with k categories. • It is assumed that the two measures are equally accurate. • is a measure of agreement that corrects for the agreement that would be expected by chance. 64 Cohen’s Kappa Table. Layout of data for computations of Cohen’s and weighted 1 1 p11 Measure 1 2 p21 (or Rater 1) . . . k pk1 Total c1 Measure 2 (or Rater 2) 2 . . k Total p12 . p1k r1 . p22 . p2k r2 . . . . . . . . . . . . pk2 . pkk rk . c2 . ck 1 . 65 Cohen’s Kappa Table. Layout of data for computations of Cohen’s and weighted Measure 2 1 2 . . k Total 1 p11 p12 . p1k r1 . Measure 1 2 p21 p22 p2k r2 . . . . . . . . . . . . . . . . k pk1 pk2 pkk rk . . Total c1 c2 ck 1 . . The observed proportion of agreement, Po, is the sum of the proportions on the diagonal: k Po pii i 1 66 Cohen’s Kappa Table. Layout of data for computations of Cohen’s and weighted Measure 2 1 2 . . k Total 1 p11 p12 . p1k r1 . Measure 1 2 p21 p22 p2k r2 . . . . . . . . . . . . . . . . k pk1 pk2 pkk rk . . Total c1 c2 ck 1 . . The expected proportion of agreement (on the diagonal), Pe, is: k Pe ri ci 1 Where ri and ci are marginal iproportions for the 1st and 2nd measure respectively. 67 Kappa Then, kappa is estimated by: Po Pe ˆ 1 Pe Which is: Observed agreement(%)-Expected agreement (%) 100% - Expected agreement (%) • 1 Pe = maximum possible nonchance agreement or 100% less the contribution of chance • Po Pe = proportion of observations that can be attributed to reliable measurement (i.e., not due to chance) • So kappa is the ratio of the number of observed nonchance agreements to the number of possible nonchance agreements 68 Pictorial of kappa statistic agreement expected by chance 0 potential improvement beyond chance 100% agreement observed agreement Kappa = % of maximum possible improvement over that expected by chance alone (kappa 0.50 here) 69 Kappa Po Pe ˆ 1 Pe • Kappa ranges from –1 (perfect disagreement) to +1 (perfect agreement) • Kappa of 0 means that: observed agreement = expected agreement 70 Reliability of categorical variables • Example 1: Agreement matrix for radiologists reading mammography for breast cancer Radiologist A Radiologist B Yes + No - Yes + No - 21 (a) 43 (b) 64 3 (c) 83 (d) 86 24 126 Overall % agreement = (a+d) / (a+b+c+d)=(21+83)/150=0.69 Po Pe 0.69 0.55 ˆ 0.31 1 Pe 1 0.55 71